A Machine Learning Approach to March Madness - Neural ...

 
CONTINUE READING
A Machine Learning Approach to March Madness∗

                                               Jared Forsyth, Andrew Wilde
                                                    CS 478, Winter 2014
                                              Department of Computer Science
                                                 Brigham Young University

                          Abstract                                      Since there are 64 games played in the NCAA tournament,
                                                                     it is nearly impossible to predict a perfect bracket. High
     The aim of this experiment was to learn which                   Point Enterprise, a morning paper from North Carolina, stated
     learning model, feature selection technique, order-             that ”you have a much greater chance of winning the lot-
     ing process, and distance criteria provided the best            tery, shooting a hole-in-one in golf or being struck by light-
     classification accuracy in predicting the outcome of            ning”. They estimated that the chances of predicting a perfect
     any NCAA Men’s Basketball Tournament match.                     bracket are 1 in 9.2 quintillion.
     Ninety-four features were selected from ESPN.com                   It became clear that developing a model that provided per-
     for each team accepted into the tournament for the              fect win/loss classification was unrealistic, so instead we fo-
     last four years. Ordering processes were tested                 cused on improving the prediction accuracy of individual
     against the baseline and the random ordering in-                games. Data was collected from a number of sources to help
     creased classification accuracy from 0.61 to 0.63.              with the learning. Certain organizations such as ESPN.com
     Random ordering was used to test a variety of                   and NCAA.com keep mounds of statistical information on
     feature reduction techniques. Random forest fea-                every team throughout the regular season. By collecting these
     ture reduction performed best and increased accu-               statistical measurements and running them through numerous
     racy from 0.63 to 0.7322 when used in conjunc-                  learning models, prediction accuracy could drastically im-
     tion with kNN and limiting the number of features               prove.
     to five. Using Manahattan distances as opposed to                  Note that the problem at hand is not classification of indi-
     Euclidean distance further increased accuracy from              vidual teams, but rather predicting the outcome of a match
     0.7322 to 0.7362.                                               between any two teams. We discuss the impact of this dis-
                                                                     tinction and our steps to deal with it in the methods section.
1   Introduction
Since 1939, the best colleges and universities across the
                                                                     2     Methods
United States have participated in a yearly tournament called        2.1    Data Source
the NCAA Men’s Basketball Championship. This basketball              We selected our data from the ESPN.com website, one of
tournament has become one of the most popular and famous             the largest and most popular sources for sports news in gen-
sporting tournaments in the United States. Millions of peo-          eral, and particularly for information about March Madness.
ple around the country participate in contests in which the          ESPN.com is listed as the third search result in Google (af-
participants make their best guesses on who will win each            ter two ncaa.com results) for the query ”march madness”.
game throughout the tournament. These types of contests              ESPN.com not only publishes the results of the March Mad-
have become so popular that more than 11 million brackets            ness tournament, but also keeps track of a vast number of
were filled out on ESPN.com in 2014. One billion dollars             statistics about each team and player that participates in the
was even being rewarded to anyone who achieved a ”perfect            NCAA tournament. Each team is associated with a unique ID
bracket” (guessed every game correctly).                             on ESPN.com, making associating match results with team
   Every game is unpredictable, and the teams that are sup-          statistics trivial.
posedly the ”better team” sometimes end up losing. This                 Though the NCAA March Madness competition has ex-
is called an upset in basketball lingo, and happens regularly        isted for many decades, ESPN changed the way that they cal-
throughout the tournament. Because of these upsets, it can           culate the statistics in 2010, leaving us with 4 years of data.
be difficult to correctly guess the winner of each game. The         Each championship contains 64 matches, yielding nearly 300
tournament can be so unpredictable that the time period over         matches for evaluation, each match representing a single in-
which the tournament runs has been termed March Madness.             stance in our machine learning problem. Also included in the
   ∗                                                                 instance are the statistics provided by ESPN.com for the two
     These match the formatting instructions of IJCAI-07. The sup-
port of IJCAI, Inc. is acknowledged.                                 competing teams, of which there are 47 per team. This re-
sults in 94 features per instance, of which several features are       Feature Name           Feature Description
derived from other features.
                                                                   DEFQ                Defensive Quotient
   We considered growing the data set by including matches         DRPG                Defensive rebounds per game
that occurred during the season preceding the tournament, but
given the specificity of the problem, we decided that the ad-      FG                  Field goal percentage
dition of such instances would not help the algorithm to gen-      FT                  Total free throws made for season
eralize on championship games. Additionally, many of the
features taken from ESPN are calculated based on the team’s        FTA                 Total free throw attempts for season
performance during the preceding season, which would not           FTM                 Free throws made for season
apply to matches that occurred during that season.
                                                                   GP                  Games played for season
                                                                   LRPI                Road and neutral game RPI

2.2    Selected Models                                             NCRP                Non-conference RPI
                                                                   NCSS                Non conference strength of schedule
                                                                   OFF                 Offensive rating
Given that the comparison classification problem is suffi-
ciently different from typical problems, we wanted to eval-        OFFQ                Offensive quotient
uate several different learning algorithms to see which would      ORPG                Average offensive rebounds per game
fare the best. All of the input features are continuous, with
a boolean output class, which fits many common algorithms          PF                  Personal fouls for season
well.                                                              PPG                 Points per game
   Accordingly, we selected Naive Bayes, kNN, Decision             PPS                 Points per shot
Tree, SVM, CN2 Induction, Random Forest, Logistic Regre-           RK                  Rank at the end of the season
sison, and Backpropagated Neural Network to form our base
set of models. Additionally, we ran all tests with a simple        RPG                 Rebounds per game
learner that output the majority class (stochastically) to form    RPI                 Rating percentage index
a baseline.
                                                                   SOS                 Strength of schedule
                                                                   ST/TO               Steals to turnover ratio
                Table 1: Feature Descriptions                      STPG                Steals per game
                                                                   TOPG                Turnovers per game
 Feature Name                Feature Description
2P                  Two point field goals made for season          These statistics refer to teams’ performances as a whole and
2PA                 Two point field goal attempts for season       do not pertain to the performance of individual players.
2PM                 Two point field goals made for season
3P                  Total three point fields for season            3     Initial Results
3PA                 Three point attempts                           For our initial results, we ran all of the learners with their
                                                                   default parameters (as supplied by the orange computation
3PM                 Three pointers made for season                 platform) using all 94 features and a feature-based order-
APG                 Assists per game                               ing scheme where the ”team1” was the team with the lower
                                                                   ”strength of schedule” score. This yielded abysmal results;
ASM                 Average scoring margin                         the best learner was kNN, which achieved 61% accuracy,
AST                 Assists for season                             barely above the baseline of 60%. All of the other learners
                                                                   tied with the baseline or did significantly worse.
AST/TO              Assists to turnover ratio
BLK                 Blocks for season                              4     Improvements
BLK/PF              Blocks to personal foul ratio                  4.1    The Problem of Comparison
BLKPG               Blocks per game                                One of the features of this machine learning task is that its
CFRP                Conference RPI                                 goal is comparison between two items (in this case teams),
                                                                   where for each team there are a number of features associated
CFSS                Conference strength of schedule                with it. The problem of classification, and specifically binary
DEF                 Defensive rating                               classification, is to determine whether an instance A is one of
                                                                   two classes, for example X or Y . What’s unique about the
situation of comparison is that each instance A is comprised        ordering resulted 90% of the instances being the same class,
of an ordered pair of ”teams” B and C. For every instance           many learners would doubtless do very well, perhaps in the
BC => X, the reversed pair CB will always Y , and vice              80-90% range. This is far less impressive, however, than a
versa.                                                              learner that achieves 60% accuracy on a data set that is split
   This adds a layer of complexity to what the learner must         50-50.
learn; for every logical input feature f , fA and fB are im-
plicitly linked and their reverse is linked to the reverse of the
outcome.
   The setup of the learners forces B and C to be ordered,
though the data set is not inherently ordered; for each game
there are just two teams and an indication of which team won.
An obvious option would be to put the winner first and the
loser second, but then all classifiers would only classify the
outcome as ”WIN”; the ordering must be something that can
be trivially determined for novel data.
   We investigated several methods of approaching this issue,
including feature-based ordering, arbitrary ordering, reverse
duplication, and random ordering.
Feature-based Ordering                                              Figure 1: Performance of Various Learners on Different Or-
Initially, we just sorted the teams based on a single feature       derings
that we thought was most useful, the ”strength of schedule”
feature that is decided by ESPN.com based on a secret algo-
rithm known only to them. This feature generally has a lot
of weight when people build their brackets by hand, and so
we tried ordering the teams such that the team with a lower
”strength of schedule” came first.
Arbitrary Ordering
Our second approach was to order the teams alphabetically,
such that the team to come ”first” was the one whose name
came first alphabetically. This was arbitrary, but not random;
it introduced a bias that was completely unrelated to the prob-
lem at hand.
Random Ordering
A third approach was to order the teams randomly – essen-
tially remove any bias from the ordering. Removing that bias        Figure 2: Performance of Various Learners against the Base-
would mean one less thing that a learner has to discover. This      line
doesn’t help them to understand that there is reversability in
the data set, however, which could potentially damage results.         The large amount of variance for a given learner on the
Data Doubling                                                       different orderings indicates that ordering does have a large
                                                                    impact on a learner’s ability to learn the problem. The various
The final method was to insert both orderings; both BC =>
                                                                    learners reacted differently to each of the orderings, as some
X and CB => Y . This would give learners the most pos-
                                                                    were better able to cope with or take advantage of the bias
sible information and hopefully also remove any bias. While
                                                                    introduced by a given scheme.
testing learners, pairs from the test set were excluded from
the test set. If BC were left in the training set when testing         For the arbitrary (alphabetical by team name) ordering, al-
on CB, the learners achieved nearly 100% accuracy.                  most all of the learners were more accurate than the baseline.
                                                                    Because the ordering was fairly uncorrelated with the learn-
Comparison of Orderings                                             ing goal, there wasn’t much of an extra bias that had to be
When tested with the default settings of eight different ma-        overcome. However, it still fared worse than the random or-
chine learning algorithms, the different ordering schemes re-       dering, which was (by definition) uncorrelated and unbiased.
sulted in significant differences in accuracy. Figure 1 shows          Feature-based ordering was least helpful; almost all of the
the absolute accuracy of each learner for each of the four dif-     learners did worse than the baseline. The ”strength of sched-
ferent ordering schemes, and Figure 2 shows the relative ac-        ule” feature was more correlated with the learning goal than
curacy compared with a baseline learner (majority) for that         any other ordering; it resulted in 60-40% win-loss split in our
ordering. Accuracy was determined using 10-fold cross vali-         data set. The majority learner therefor achieved 60% accu-
dation.                                                             racy, making all others comparatively much worse.
   In this comparison, difference from the baseline is a more          Random ordering was most effective across the board, both
useful metric than absolute performance. In a case where the        in terms of absolute accuracy and in terms of improvement
over the baseline. Given that none of the algorithms used are      duction achieved. The kNN algorithm was boosted by 10%
designed to take advantage of ordering bias in a comparison        when run on only the top 5 features from the random forest
instance, the best method was to just remove the bias.             weights—a 95% reduction in feature space size. All except
   Data doubling was surprisingly ineffective. Even though         logistic regression and SVM achieved similar improvements
there were twice as many instances for a learner to learn from     in accuracy. As was the case with SVM reduction, the Ran-
and the ordering bias was effectively neutralized, the prob-       dom Forest reduction resulted in a significant improvement
lem was apparently just as hard to learn as when there was         for the Random Forest algorithm.
ordering bias.

4.2   Feature Reduction
From an initial feature space of 94 dimensions, we tested
several different methods of reduction, in which we took the
top n attributes as ranked by four different algorithms: Gain
Ratio, ReliefF, Linear SVM Weights, and Random Forests.
For each algorithm we tested taking different numbers of at-
tributes (always the top n) for each of the learners. Figure
3 shows the best accuracy achieved by each learner for each
reduction technique, for n < 45.

                                                                         Figure 4: Benefit of Reduction to Individual Learners

                                                                   4.3     Model Tweaks
                                                                   The kNN learner achieved the greatest accuracy after apply-
                                                                   ing feature reduction, so we chose to focus on it for more
                                                                   granular parameter adjustment. We experimented with two
                                                                   different distance metrics, Euclidean and Manhattan, at vary-
                                                                   ing values of k to isolate the ideal parameters. Results are
                                                                   displayed in Figure 5. The greatest accuracy (73.62%) was
                                                                   reached by using the Manhattan distance metric and a k value
                                                                   of 130.
        Figure 3: Comparison of Reduction Methods

   ReliefF scoring proved to be the least helpful, and was the
only method that actually decreased accuracy when the num-
ber of features was reduced; the other three methods managed
to achieve significant gains over the full feature set.
   Using Gain Ratio to reduce feature size resulted in modest
accuracy improvements for over half of the learners. How-
ever, it was only able to reduce the feature set size down to
about 15 before all the learners began to lose accuracy dra-
matically.
   The two reduction methods that involved more complex
computation resulted in the most impressive accuracy gains.
They also succeeded in reducing the number of features by              Figure 5: Nearest Neighbor Accuracies by # Neighbors
over 80%.
   Reduction using the Support Vector Machine resulted in
dramatic improvements for several learners. The SVM                5     Results
learner improved by 8% and the neural network and logistic         Numerous tests were performed altering the ordering se-
regression algorithms both improved by 6%. In each case, the       quence, training model, features, and type of distance (Eu-
feature space was reduced from 94 features to between 5 and        clidean vs. Manhattan). The data was trained using SVM,
10, dramatically speeding up learning time as well. It is inter-   CN2 rules, neural network, classification tree, logistic regres-
esting that the SVM learner improved when using SVM pre-           sion, naive bayes, random forest, k-nearest neighbor, or sim-
processing. The second learner was apparently able to benefit      ple majority. The features were selected according to those
from the work done by the previous one, while at the same          that provided the highest SVM weights. In other words, if
time exploring new territory to achieve better accuracy.           the data set was training using 10 features, then the 10 fea-
   Random forest feature reduction proved to be the most suc-      tures with the highest SVM weights were used. Increasing
cessful, both in terms of accuracy and in the amount of re-        classification accuracy (CA) was our primary goal.
5.1   Ordering                                                    Model                           Top N Features
Each ordering technique was run with each learning model.                               All       15        5         4
The results can be seen on table 2.                               SVM                   0.6195    0.6452    0.6228    0.6226
                                                                  CN2 rules             0.6417    0.6687    0.6835    0.6905
                      Arbitrary   Random       Feature   Double   Random Forest         0.6189    0.6903    0.6719    0.6571
                                                                  Neural Network        0.6379    0.6493    0.6942    0.6946
SVM                   0.6047      0.6195       0.5813    0.6071
                                                                  Naive Bayes           0.6313    0.6755    0.6909    0.6835
CN2 rules             0.5734      0.6417       0.555     0.5854
                                                                  Logistic regression   0.6533    0.6379    0.6187    0.6187
Random Forest         0.5249      0.6189       0.5171    0.5
                                                                  kNN                   0.6312    0.6872    0.7322    0.7024
Neural Network        0.6043      0.6379       0.5736    0.5783
                                                                  Classification Tree   0.6229    0.5774    0.6756    0.5936
Naive Bayes           0.6201      0.6313       0.6077    0.5613
Logistic regression   0.5852      0.6533       0.6004    0.6034
Majority              0.5246      0.517        0.6001    0.5528 Table 4: Features were selected by selecting those features
kNN                   0.5816      0.6312       0.6117    0.5411 most important in a random forest
Classification Tree   0.5506      0.6229       0.55      0.609
                                                                Model                          Top N Features
        Table 2: Ordering influence on training models                                  All       65        15        5
                                                                  SVM                   0.6195    0.6085    0.5991    0.5651
   Random ordering gave us our best results for every training    CN2 rules             0.6417    0.5813    0.5969    0.4828
model. The classification accuracy increased from the base-       Random Forest         0.6189    0.616     0.5665    0.5132
line of 0.61 to 0.63. Other ordering techniques introduced a      Neural Network        0.6379    0.6303    0.6305    0.5694
type of bias that random ordering did not introduce. These        Naive Bayes           0.6313    0.6496    0.6074    0.566
results were attained without implementing any feature re-        Logistic regression   0.6533    0.6528    0.6377    0.5991
duction techniques.                                               kNN                   0.6312    0.6199    0.5701    0.5963
                                                                  Classification Tree   0.6229    0.5595    0.5661    0.5131
5.2   Feature Reduction
SVM                                                                     Table 5: Features selected using ReliefF scoring
Features were chosen according to their SVM weights. If
only ten features were selected, the top ten features with the    Gain Ratio
highest SVM weights were used. Random ordering was used           Calculates features that provide the most gain, but does not
in creating each instance.                                        take combination of features into account. Table 8 shows fea-
                                                                  ture selection using gain ratio.
Model                             Top N Features
                      All      35      10       9        5        Model                           Top N Features
SVM                   0.619    0.679   0.702    0.698    0.588                          All       35        15        5
CN2 rules             0.641    0.663   0.562    0.592    0.569    SVM                   0.6195    0.6566    0.6298    0.6148
Random Forest         0.618    0.660   0.611    0.641    0.551    CN2 rules             0.6417    0.6342    0.6567    0.5698
Neural Network        0.637    0.667   0.687    0.706    0.641    Random Forest         0.6189    0.6386    0.6835    0.5922
Naive Bayes           0.631    0.645   0.634    0.623    0.6      Neural Network        0.6379    0.6224    0.6684    0.615
Logistic regression   0.653    0.668   0.691    0.720    0.607    Naive Bayes           0.6313    0.6642    0.6528    0.6225
kNN                   0.631    0.672   0.656    0.657    0.603    Logistic regression   0.6533    0.6349    0.6115    0.6074
Classification Tree   0.622    0.555   0.547    0.570    0.572    kNN                   0.6312    0.6952    0.6989    0.6679
                                                                  Classification Tree   0.6229    0.6342    0.5671    0.5775
Table 3: Features were reduced using the best features ac-
cording to SVM weights.                                                    Table 6: Features selected using gain ratio.

Random Forest                                                     Feature Reduction Summary
Features were chosen by constructing a large number of de-        Random forest provided the best results when used in con-
cision trees and choosing the features that averaged the most     junction with kNN and limiting the number of features used
importance among the decision trees. These results can be         in the training to five. The classification accuracy increased
viewed in Table 4.                                                to from 0.63 to 0.7322, adding more than 10% of additional
                                                                  accuracy.
ReliefF
Instances are chosen at random and changes the weights of         5.3   Euclidean Distance vs. Manhattan Distance
feature relevance according to its nearest neighbor. Table 7      The results on Table 9 show the differences between us-
shows the results using ReliefF scoring.                          ing Euclidean Distance and Manhattan Distance on k-nearest
neighbor training on a randomly ordered instances and limit-        6.4    Random Forest and Combination of Features
ing the number of features used in training to five.                       Theory
                                                                    Using random forest takes into account how features inter-
# Neighbors     Euclidean     Manhattan                             act one with another. This allowed us to use those features
                                                                    that provided the best results when used in collaboration with
60              0.7174        0.6984                                other features. Thus, we were better able to compensate for
80              0.7209        0.7098                                the importance of a combination of features and use that as
100             0.7322        0.7211                                part of the training process.
120             0.7171        0.7286
130             0.7095        0.7362                                6.5    Perfect Bracket Probabilities
140             0.7020        0.7246
                                                                    Even with large amounts of statistical information, there are
                                                                    some parts of a basketball game you simply cannot predict.
Table 7: Nearest Neighbor Distance Algorithms and K values          Injuries cannot be forecast, and the mental state of players
                                                                    playing the game cannot be observed. However, these things
   Using a Manhattan distance and the nearest 130 neighbors         can drastically affect the outcome of a basketball match. Per-
increased our previous best result from 0.7322 to 0.7362.           fect brackets will therefore continue to be impossible to cal-
                                                                    culate, no matter how much data is provided to the learning
                                                                    models.
6     Conclusion
The results of our experiments show that random ordering            7     Future Work
used in conjunction with k-nearest neighbor, Manhattan dis-
                                                                    We believe that our classification accuracy can improve to
tance, and random forest as the feature reduction algorithm
                                                                    85% if the proper features are used in training. Vital statistical
provide the best classification accuracy. Our best classifica-
                                                                    information we were unable to gather include the following:
tion accuracy, using each of these algorithms, was 0.7362.
The following sections describe possible causes about why               • Individual statistics for players on each team
these features provided the best results among all experiments          • Venue of the match including the distance
that were tested.
                                                                        • Importance of a match-up
6.1    Style Theory                                                     • Win/loss record for the last five, ten, and fifteen games
In basketball, it is believed that certain teams match-up better        • Average number of players who play consistently
against certain teams. For example, one team may be ex-             All of this data is expected to play a role in how teams per-
tremely quick while another team is extremely tall. If the first    form. However, this type of data was out of our scope to ac-
team can alter the game so that it is played at a fast pace, then   cumulate. Organizations with more access to this type of data
the game style is in their favor. Essentially, specialties on a     will be able to study whether these features have a profound
basketball team translate to success against opponents with         effect on the success of a basketball team.
certain characteristics.                                               This research would also benefit greatly with more data.
                                                                    Since only the statistics for the last four tournaments were ac-
6.2    K-nearest Neighbor & Style Theory                            cessible in a consistent format, our training models risked the
                                                                    danger of overfit. We would recommend that much more data
K-nearest neighbor provided the best results in comparison
                                                                    be gathered to help avoid overfit. However, since basketball
to the other training models most likely due to this theory.
                                                                    continually changes throughout the years, we recommend not
The data compares itself to the instances that provide similar
                                                                    use data from more than ten years back.
data. In other words, quicker teams that play taller teams will
                                                                       We would also recommend experimenting with using sev-
compare themselves to other match-ups that involved quicker
                                                                    eral feature reduction techniques together as opposed to using
teams playing taller teams. The consistencies of the match-
                                                                    the reduction techniques one by one.
ups and their associated outcomes will mean more to an in-
stance that shares those characteristics.

6.3    Combination of Features Theory
Some features, when combined, prove to be more beneficial
to the success of a team than others. For example, a team that
has a low number of steals per game, shoots a large number
of field goals, and shoots a high field goal percentage will
have more success than a team with a combination of large
number of steals, high field goal percentage, and low number
of shots taken each game. Learning models that take into ac-
count the combination of features when classifying instances
will perform better than those that do not.
You can also read